Aerial photograph of the Immaculate Conception Chapel

Classification in LiDAR

Object Segmentation and Classification from LiDAR Point Clouds

LiDAR is a remote sensing method which produces precise 3D point clouds which consist of millions of geo-spatially located data points and possible color information. Because of the nature of LiDAR point clouds it can often be difficult for analysts to accurately and efficiently recognize and categorize objects. The overall goal of this project is automatic object segmentation and classification of LiDAR point clouds captured by aerial systems.

Example 

The algorithm consists of two parts. The first part is object segmentation. The point cloud information is encoded as a grid of vertical 3D strips where each strip is a histogram of the points that fall within the space. The size of these strips can be dynamically determined by examining the distribution between neighboring points within the strip.  From this strip histogram representation a set of features, including the local ground and relative variance within the strip can be calculated. Using the standard deviation and Euclidean distance between a strip and its neighbors we can identify seeds which can be used for region growing based segmentation.

Lidar Flowchart

The second part of the algorithm is classification of the segmented objects. Once individual objects are segmented, a set of geometrical features of each object region are calculated, including 3D form factor, texture, size/area of the region, orientation of the surface, average point density, average curvature and average surface smoothness These features are used as inputs to a support vector machine (SVM), which is used to classify individual objects in the scene.

Strip Histogram Point Cloud

Image Demonstrations

Region Growing Region Growing